Integrating Cache Related Preemption Delay Analysis into EDF Scheduling

Transcription

1 Integrating Cache Related Preemption Delay Analysis into EDF Scheduling Will Lunniss 1 Sebastian Altmeyer 2 Claire Maiza 3 Robert I. Davis 1 1 Real-Time Systems Research Group, University of York, UK {wl510, rob.davis}@york.ac.uk 2 Department of Computer Science, Saarland University, Germany altmeyer@cs.uni-sb.de 2 Verimag, Grenoble INP, Grenoble, France Claire.maiza@imag.fr 1

2 Outline EDF Scheduling CRPD Integrating CRPD analysis into EDF Comparison with existing approach Improved CRPD analysis Case study Synthetic taskset evaluation Conclusions 2

3 Earliest Deadline First (EDF) It is dynamic scheduling algorithm Schedules the job of the task with the earliest absolute deadline first Proven to be optimal by Dertouzos on a single core processor 3

4 Determining which job should run 4

5 Determining which job should run If two jobs have the same absolute deadline We assume that the job with the lowest task index is chosen E.g. τ 2 pre-empts τ 3 in the above example 4

6 Determining which job should run If two jobs have the same absolute deadline Ensures that two tasks cannot pre-empt each other Ensures that after a pre-emption, the task that was pre-empted last is resumed first E.g τ 2 is resumed at t = 7 5

7 Determining which job should run If two jobs have the same absolute deadline Ensures that two tasks cannot pre-empt each other Ensures that after a pre-emption, the task that was pre-empted last is resumed first E.g τ 2 is resumed at t = 7, rather than τ 3 5

8 Determining which job should run Also applies for jobs with the same relative deadline and release time E.g. τ 2 is resumed at t = 3, rather than τ 3 starting 6

9 Determining which job should run Also applies for jobs with the same relative deadline and release time E.g. τ 2 is resumed at t = 3, rather than τ 3 starting 6

10 Determining which job should run Also applies for jobs with the same relative deadline and release time E.g. τ 2 is resumed at t = 3, rather than τ 3 starting 7

11 Schedulability test If all tasks have implicit deadlines (D i =T i ), schedulability test is 8

12 Schedulability test If all tasks have implicit deadlines (D i =T i ), schedulability test is U 1 8

13 Schedulability test If all tasks have implicit deadlines (D i =T i ), schedulability test is U 1 If D i T i then the test is still necessary, but is no longer sufficient 8

14 Schedulability test If all tasks have implicit deadlines (D i =T i ), schedulability test is U 1 If D i T i then the test is still necessary, but is no longer sufficient Need to do another test 8

15 Processor demand bound function h( t) i1 t D max0,1 Ti i C i 9

16 Processor demand bound function h( t) i1 t D max0,1 Ti i C i Sum over each task 9

17 Processor demand bound function h( t) i1 t D max0,1 Ti i C i Sum over each task the number of jobs a task has which are released and have their deadlines in the interval t 9

18 Processor demand bound function h( t) i1 t D max0,1 Ti i C i Sum over each task the number of jobs a task has which are released and have their deadlines in the interval t multiplied by the tasks execution time 9

19 Processor demand bound function h( t) i1 t D max0,1 Ti i C i Sum over each task the number of jobs a task has which are released and have their deadlines in the interval t multiplied by the tasks execution time S. K. Baruah, L. E. Rosier, and R. R. Howell, "Algorithms and Complexity Concerning the Preemptive Scheduling of Periodic Real-Time Tasks on One Processor," Real-Time Systems, vol. 2, no. 4, pp ,

20 Schedulability test A taskset is schedulable iff h(t) t for all values of t The execution time requirement must be less than or equal to the available time 10

21 Schedulability test A taskset is schedulable iff h(t) t for all values of t The execution time requirement must be less than or equal to the available time h(t) can only change when t is equal to an absolute deadline Bound the maximum value of t, L, using either Hyper-period: Least common multiple of task periods Synchronous busy period 10

22 Schedulability test There are still a large number of values for t that need to be checked 11

23 Schedulability test There are still a large number of values for t that need to be checked Can be reduced by using the Quick convergence Processor-demand Analysis (QPA) algorithm by Zhang and Burns Starts with a value of t close to L Iterates back towards 0 checking a significantly smaller number of values 11

24 Schedulability test There are still a large number of values for t that need to be checked Can be reduced by using the Quick convergence Processor-demand Analysis (QPA) algorithm by Zhang and Burns Starts with a value of t close to L Iterates back towards 0 checking a significantly smaller number of values F. Zhang and A. Burns, "Schedulability Analysis for Real-Time Systems with EDF Scheduling," IEEE Transactions on Computers, vol. 58, no. 9, pp , September

25 Pre-emptions and Cache Related Pre-empt Delays (CRPD) 12

26 Pre-emptions and Cache Related Pre-empt Delays (CRPD) Pre-empting task can evict blocks belonging to the pre-empted task CRPD are introduced when the pre-empted task has to reload some of those evicted cache blocks after resuming 12

27 Cache block categorisation Evicting Cache Blocks (ECBs) Loaded into cache and can therefore evict other blocks 13

28 Cache block categorisation Evicting Cache Blocks (ECBs) Loaded into cache and can therefore evict other blocks Useful Cache Blocks (UCBs) Reused once they have been loaded into cache before potentially being evict by the task If evicted by another task, they may have to be reloaded which introduces CRPD UCBs are always ECBs 13

29 Cache block categorisation Example block classification 14

30 Cache block categorisation Example block classification Instructions inside loops are often UCBs as they get reused 14

31 CRPD analysis Need to calculate the number of blocks evicted during a pre-emption that must be reloaded Multiply by the cost to reload each block, BRT 15

32 CRPD analysis Need to calculate the number of blocks evicted during a pre-emption that must be reloaded Multiply by the cost to reload each block, BRT Could take a simple approach and assume every block evicted by a pre-empting task must be re-loaded e.g. BRT t, j ECB j 15

33 CRPD analysis Adapted a number of approaches for FP to work with EDF Defined: the sets of tasks which can/cannot pre-empt each other how often these pre-emptions can occur within the interval t 16

34 CRPD analysis Adapted a number of approaches for FP to work with EDF Defined: the sets of tasks which can/cannot pre-empt each other how often these pre-emptions can occur within the interval t Then include the CRPD into the h(t) calculation 16

35 Integrating CRPD analysis into the h(t) calculation n t D j h( t) max0,1 C j t j1 Tj, j 17

36 Integrating CRPD analysis into the h(t) calculation Calculate the CRPD caused by one job of task τ j in the interval t n t D j h( t) max0,1 C j t j1 Tj, j 17

37 Integrating CRPD analysis into the h(t) calculation Calculate the CRPD caused by one job of task τ j in the interval t n t D j h( t) max0,1 C j t j1 Tj, j Then add it to the execution time of that job of task τ j 17

38 Effect of CRPD on task utilisation and h(t) calculation The analysis effectively increases the execution time of a task by the CRPD it causes Need to account for this when calculating the utilisation of a task and taskset Also need to use this when calculating the upper bound of t used for calculating h(t) 18

39 Set of pre-empting tasks Based on the tasks relative deadline Assume that any task τ j with a relative deadline D j < D i can pre-empt task τ i 19

40 Set of pre-empting tasks Based on the tasks relative deadline Assume that any task τ j with a relative deadline D j < D i can pre-empt task τ i The set of tasks that can pre-empt task τ i is: 19

41 Set of pre-empting tasks Based on the tasks relative deadline Assume that any task τ j with a relative deadline D j < D i can pre-empt task τ i The set of tasks that can pre-empt task τ i is: hp( i) { j D j Di } 19

42 Set of pre-empted tasks Task τ j can pre-empt any tasks whose relative deadline is greater than it s relative deadline 20

43 Set of pre-empted tasks Task τ j can pre-empt any tasks whose relative deadline is greater than it s relative deadline Can exclude all tasks whose relative deadlines are greater than t They do not need to be included when calculating h(t), see paper for proof 20

44 Set of pre-empted tasks Task τ j can pre-empt any tasks whose relative deadline is greater than it s relative deadline Can exclude all tasks whose relative deadlines are greater than t They do not need to be included when calculating h(t), see paper for proof aff t, j i t Di D j 20

45 UCB-Union Based on approach by Tan and Mooney ucbu t, j BRT UCBk kaff( t, j) ECB j 21

46 UCB-Union Based on approach by Tan and Mooney ucbu t, j BRT UCBk kaff( t, j) ECB j 21

47 UCB-Union Based on approach by Tan and Mooney ucbu t, j BRT UCBk kaff( t, j) ECB j Calculate the union of the UCBs of all tasks that: can be evicted by the pre-empting task τ j have a job with a release time and absolute deadline within the interval t 21

48 UCB-Union Based on approach by Tan and Mooney ucbu t, j BRT UCBk kaff( t, j) ECB j Calculate the union of the UCBs of all tasks that: can be evicted by the pre-empting task τ j have a job with a release time and absolute deadline within the interval t intersected with the ECBs of the pre-empting task 21

49 UCB-Union Based on approach by Tan and Mooney ucbu t, j BRT UCBk kaff( t, j) ECB j Calculate the union of the UCBs of all tasks that: can be evicted by the pre-empting task τ j have a job with a release time and absolute deadline within the interval t intersected with the ECBs of the pre-empting task Y. Tan and V. Mooney, "Timing Analysis for Preemptive Multitasking Real-Time Systems with Caches," ACM Transactions on Embedded Computing Systems (TECS), vol. 6, no. 1, February

50 ECB-Union Based on the approach by Altmeyer et al. UCB ecbu t, j BRT max k kaff( t, j) hhp( j) { j} ECB h 22

51 ECB-Union Based on the approach by Altmeyer et al. UCB ecbu t, j BRT max k kaff( t, j) hhp( j) { j} ECB h Assume that task τ j has already been pre-empted. Include the union of ECBs belonging to all tasks that can pre-empt it 22

52 ECB-Union Based on the approach by Altmeyer et al. UCB ecbu t, j BRT max k kaff( t, j) hhp( j) { j} ECB h Assume that task τ j has already been pre-empted. Include the union of ECBs belonging to all tasks that can pre-empt it 22

53 ECB-Union Based on the approach by Altmeyer et al. UCB ecbu t, j BRT max k kaff( t, j) hhp( j) { j} ECB h Calculate the maximum number of UCBs that may need to be reloaded by any task that is directly pre-empted by task τ j Assume that task τ j has already been pre-empted. Include the union of ECBs belonging to all tasks that can pre-empt it 22

54 ECB-Union Based on the approach by Altmeyer et al. UCB ecbu t, j BRT max k kaff( t, j) hhp( j) { j} ECB h Calculate the maximum number of UCBs that may need to be reloaded by any task that is directly pre-empted by task τ j Assume that task τ j has already been pre-empted. Include the union of ECBs belonging to all tasks that can pre-empt it S. Altmeyer, R.I. Davis, and C. Maiza, "Cache Related Pre-emption Delay Aware Response Time Analysis for Fixed Priority Pre-emptive Systems," in Proceedings of the 32nd IEEE Real-Time Systems Symposium (RTSS), Vienna, Austria, 2011, pp

55 Existing approach In 2007, Ju et al. presented an approach We refer to it as the JCR approach after their initials jcr i BRT P j jhp( i) ( D i ) UCB i ECB j 23

56 Existing approach In 2007, Ju et al. presented an approach We refer to it as the JCR approach after their initials jcr i BRT P j jhp( i) ( D i ) UCB i ECB j Sum for every task τ j that can pre-empt task τ i 23

57 Existing approach In 2007, Ju et al. presented an approach We refer to it as the JCR approach after their initials jcr i BRT P j jhp( i) ( D i ) UCB i ECB j Sum for every task τ j that can pre-empt task τ i the number of times task τ j can pre-empt a single job of task τ i 23

58 Existing approach In 2007, Ju et al. presented an approach We refer to it as the JCR approach after their initials jcr i BRT P j jhp( i) ( D i ) UCB i ECB j Sum for every task τ j that can pre-empt task τ i the number of times task τ j can pre-empt a single job of task τ i multiplied with the number of task τ i UCBs that could be evicted task τ j ECBs 23

59 Existing approach In 2007, Ju et al. presented an approach We refer to it as the JCR approach after their initials jcr i BRT P j jhp( i) ( D i ) UCB i ECB j Sum for every task τ j that can pre-empt task τ i the number of times task τ j can pre-empt a single job of task τ i multiplied with the number of task τ i UCBs that could be evicted task τ j ECBs Ju, S. Chakraborty, and A. Roychoudhury, "Accounting for Cache-Related Preemption Delay in Dynamic Priority Schedulability Analysis," in Design, Automation and Test in Europe Conference and Exposition (DATE), Nice, France, 2007, pp

60 Existing approach Can be pessimistic for nested pre-emptions Calculates the cost between each pair of tasks 24

61 Existing approach Can be pessimistic for nested pre-emptions Calculates the cost between each pair of tasks If pre-empting tasks have ECBs located in the same cache sets, they will be counted multiple times More likely when there the cache utilisation is high 24

62 Existing approach Can be pessimistic for nested pre-emptions Calculates the cost between each pair of tasks If pre-empting tasks have ECBs located in the same cache sets, they will be counted multiple times More likely when there the cache utilisation is high It is incomparable to the approaches we have presented so far 24

63 Improved CRPD analysis The UCB-Union and ECB-Union approach can be pessimistic They assume intermediate tasks are pre-empted the same number of times as the pre-empted task 25

64 Improved CRPD analysis E.g. the cost of τ 2 pre-empting task τ 3 is counted three times rather than once 26

65 Improved CRPD analysis E.g. the cost of τ 2 pre-empting task τ 3 is counted three times rather than once 26

66 Multiset approaches ECB-Union Multiset and UCB-Union Multiset Factor in the number of times that intermediate tasks pre-empt the pre-empted task to tighten the bound See paper for details 27

67 Comparison of approaches 28

68 Comparison of approaches 28

69 Comparison of approaches 28

70 Comparison of approaches 28

71 Comparison of approaches 28

72 Case study Based on a code from the Mälardalen benchmark suite to create a 15 task taskset Setup to model an ARM7 10MHz CPU 2KB direct-mapped instruction cache Line size of 8 Bytes, 4 Byte instructions, 256 cache sets Block reload time of 8μs 29

73 Case study Breakdown utilisation No pre-emption cost 1 Combined Multiset ECB-Union Multiset UCB-Union Multiset ECB-Union UCB-Union UCB-Only ECB-Only JCR

74 Synthetic tasksets 10 tasks per taskset 10,000 tasksets for baseline evaluation 512 cache sets Cache utilisation of 5 Maximum UCB percentage of 30% 31

75 Baseline evaluation 32

76 Baseline evaluation 32

77 Baseline evaluation Combined multiset approach outperforms all other approaches Including the existing approach 32

78 Baseline evaluation Weighted Schedulability No pre-emption cost 1 Combined Multiset ECB-Union Multiset UCB-Union Multiset ECB-Union UCB-Union UCB-Only ECB-Only JCR

79 Baseline evaluation Weighted Schedulability No pre-emption cost 1 Combined Multiset ECB-Union Multiset UCB-Union Multiset ECB-Union UCB-Union UCB-Only ECB-Only JCR

80 Varying Cache Utilisation 34

81 Varying Cache Utilisation 34

82 Varying Cache Utilisation 34

83 Varying Maximum UCB Percentage 35

84 Varying Maximum UCB Percentage 35

85 Varying Maximum UCB Percentage 35

86 Varying Number of Tasks 36

87 Conclusion Presented new CRPD aware analysis for EDF Combined multiset approach dominates the existing approach by Ju et al. Confirmed via evaluation/simulation Detailed study shows the strengths and weaknesses of the different approaches We plan to investigate which is better, FP or EDF, when taking into account CRPD 37

88 Thank you for listening Any Questions? 38